Migrating from codex-harnesses

If you already use codex-harnesses and want the same workflow under Claude Code, here is the mapping.

File mapping

codex-harnesses	claude-harnesses
AGENTS.md	CLAUDE.md
skills/<name>/SKILL.md	same path; format identical
hooks/<name>/hook.py	plugins/safety-pack/scripts/<name>.py (rewritten for Claude's JSON envelope)
policies/<name>.yaml	settings/<name>.json (rewritten for Claude's permissions schema)
templates/agents/<flavor>/AGENTS.md	templates/claude-md/<flavor>/CLAUDE.md
examples/<flavor>-project/	examples/<flavor>-project/
scripts/install.sh	scripts/install.sh (added --pack)
scripts/verify.sh	scripts/verify.sh (env var rename only)

Conceptual differences

• Hooks are not stand-alone scripts in Claude Code; they are entries in settings.json. claude-harnesses ships them inside plugins so installation wires them automatically.
• Permissions live in settings.json, not in YAML policies. The translation is mechanical: mode.approval → defaultMode, guards.block_* → permissions.deny entries, git.allow_force_push: false → Bash(git push --force*) in deny.
• Skills are first-class in Claude Code and installable via /plugin, gh skill install, or npx skills add. codex-harnesses copies skill files; claude-harnesses also offers plugin-based install.

Side-by-side adoption

You can run both at once. The skills directory (skills/<name>/SKILL.md) is shared in spirit between the two repos; behaviors and verification scripts overlap. Hooks differ because the lifecycle primitives differ.